projects/linux·blob v5.14 · refs

2113 lines · 54.3 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/userfaultfd.c
4 *
5 * Copyright (C) 2007 Davide Libenzi <[email protected]>
6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc.
8 *
9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing).
11 */
12
13#include <linux/list.h>
14#include <linux/hashtable.h>
15#include <linux/sched/signal.h>
16#include <linux/sched/mm.h>
17#include <linux/mm.h>
18#include <linux/mmu_notifier.h>
19#include <linux/poll.h>
20#include <linux/slab.h>
21#include <linux/seq_file.h>
22#include <linux/file.h>
23#include <linux/bug.h>
24#include <linux/anon_inodes.h>
25#include <linux/syscalls.h>
26#include <linux/userfaultfd_k.h>
27#include <linux/mempolicy.h>
28#include <linux/ioctl.h>
29#include <linux/security.h>
30#include <linux/hugetlb.h>
31
32int sysctl_unprivileged_userfaultfd __read_mostly;
33
34static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
35
36enum userfaultfd_state {
37 UFFD_STATE_WAIT_API,
38 UFFD_STATE_RUNNING,
39};
40
41/*
42 * Start with fault_pending_wqh and fault_wqh so they're more likely
43 * to be in the same cacheline.
44 *
45 * Locking order:
46 * fd_wqh.lock
47 * fault_pending_wqh.lock
48 * fault_wqh.lock
49 * event_wqh.lock
50 *
51 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
52 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
53 * also taken in IRQ context.
54 */
55struct userfaultfd_ctx {
56 /* waitqueue head for the pending (i.e. not read) userfaults */
57 wait_queue_head_t fault_pending_wqh;
58 /* waitqueue head for the userfaults */
59 wait_queue_head_t fault_wqh;
60 /* waitqueue head for the pseudo fd to wakeup poll/read */
61 wait_queue_head_t fd_wqh;
62 /* waitqueue head for events */
63 wait_queue_head_t event_wqh;
64 /* a refile sequence protected by fault_pending_wqh lock */
65 seqcount_spinlock_t refile_seq;
66 /* pseudo fd refcounting */
67 refcount_t refcount;
68 /* userfaultfd syscall flags */
69 unsigned int flags;
70 /* features requested from the userspace */
71 unsigned int features;
72 /* state machine */
73 enum userfaultfd_state state;
74 /* released */
75 bool released;
76 /* memory mappings are changing because of non-cooperative event */
77 bool mmap_changing;
78 /* mm with one ore more vmas attached to this userfaultfd_ctx */
79 struct mm_struct *mm;
80};
81
82struct userfaultfd_fork_ctx {
83 struct userfaultfd_ctx *orig;
84 struct userfaultfd_ctx *new;
85 struct list_head list;
86};
87
88struct userfaultfd_unmap_ctx {
89 struct userfaultfd_ctx *ctx;
90 unsigned long start;
91 unsigned long end;
92 struct list_head list;
93};
94
95struct userfaultfd_wait_queue {
96 struct uffd_msg msg;
97 wait_queue_entry_t wq;
98 struct userfaultfd_ctx *ctx;
99 bool waken;
100};
101
102struct userfaultfd_wake_range {
103 unsigned long start;
104 unsigned long len;
105};
106
107static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
108 int wake_flags, void *key)
109{
110 struct userfaultfd_wake_range *range = key;
111 int ret;
112 struct userfaultfd_wait_queue *uwq;
113 unsigned long start, len;
114
115 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
116 ret = 0;
117 /* len == 0 means wake all */
118 start = range->start;
119 len = range->len;
120 if (len && (start > uwq->msg.arg.pagefault.address ||
121 start + len <= uwq->msg.arg.pagefault.address))
122 goto out;
123 WRITE_ONCE(uwq->waken, true);
124 /*
125 * The Program-Order guarantees provided by the scheduler
126 * ensure uwq->waken is visible before the task is woken.
127 */
128 ret = wake_up_state(wq->private, mode);
129 if (ret) {
130 /*
131 * Wake only once, autoremove behavior.
132 *
133 * After the effect of list_del_init is visible to the other
134 * CPUs, the waitqueue may disappear from under us, see the
135 * !list_empty_careful() in handle_userfault().
136 *
137 * try_to_wake_up() has an implicit smp_mb(), and the
138 * wq->private is read before calling the extern function
139 * "wake_up_state" (which in turns calls try_to_wake_up).
140 */
141 list_del_init(&wq->entry);
142 }
143out:
144 return ret;
145}
146
147/**
148 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
149 * context.
150 * @ctx: [in] Pointer to the userfaultfd context.
151 */
152static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
153{
154 refcount_inc(&ctx->refcount);
155}
156
157/**
158 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
159 * context.
160 * @ctx: [in] Pointer to userfaultfd context.
161 *
162 * The userfaultfd context reference must have been previously acquired either
163 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
164 */
165static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
166{
167 if (refcount_dec_and_test(&ctx->refcount)) {
168 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
169 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
170 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
171 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
172 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
173 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
174 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
175 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
176 mmdrop(ctx->mm);
177 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
178 }
179}
180
181static inline void msg_init(struct uffd_msg *msg)
182{
183 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
184 /*
185 * Must use memset to zero out the paddings or kernel data is
186 * leaked to userland.
187 */
188 memset(msg, 0, sizeof(struct uffd_msg));
189}
190
191static inline struct uffd_msg userfault_msg(unsigned long address,
192 unsigned int flags,
193 unsigned long reason,
194 unsigned int features)
195{
196 struct uffd_msg msg;
197 msg_init(&msg);
198 msg.event = UFFD_EVENT_PAGEFAULT;
199 msg.arg.pagefault.address = address;
200 /*
201 * These flags indicate why the userfault occurred:
202 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
203 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
204 * - Neither of these flags being set indicates a MISSING fault.
205 *
206 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
207 * fault. Otherwise, it was a read fault.
208 */
209 if (flags & FAULT_FLAG_WRITE)
210 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
211 if (reason & VM_UFFD_WP)
212 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
213 if (reason & VM_UFFD_MINOR)
214 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
215 if (features & UFFD_FEATURE_THREAD_ID)
216 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
217 return msg;
218}
219
220#ifdef CONFIG_HUGETLB_PAGE
221/*
222 * Same functionality as userfaultfd_must_wait below with modifications for
223 * hugepmd ranges.
224 */
225static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
226 struct vm_area_struct *vma,
227 unsigned long address,
228 unsigned long flags,
229 unsigned long reason)
230{
231 struct mm_struct *mm = ctx->mm;
232 pte_t *ptep, pte;
233 bool ret = true;
234
235 mmap_assert_locked(mm);
236
237 ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
238
239 if (!ptep)
240 goto out;
241
242 ret = false;
243 pte = huge_ptep_get(ptep);
244
245 /*
246 * Lockless access: we're in a wait_event so it's ok if it
247 * changes under us.
248 */
249 if (huge_pte_none(pte))
250 ret = true;
251 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
252 ret = true;
253out:
254 return ret;
255}
256#else
257static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
258 struct vm_area_struct *vma,
259 unsigned long address,
260 unsigned long flags,
261 unsigned long reason)
262{
263 return false; /* should never get here */
264}
265#endif /* CONFIG_HUGETLB_PAGE */
266
267/*
268 * Verify the pagetables are still not ok after having reigstered into
269 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
270 * userfault that has already been resolved, if userfaultfd_read and
271 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
272 * threads.
273 */
274static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
275 unsigned long address,
276 unsigned long flags,
277 unsigned long reason)
278{
279 struct mm_struct *mm = ctx->mm;
280 pgd_t *pgd;
281 p4d_t *p4d;
282 pud_t *pud;
283 pmd_t *pmd, _pmd;
284 pte_t *pte;
285 bool ret = true;
286
287 mmap_assert_locked(mm);
288
289 pgd = pgd_offset(mm, address);
290 if (!pgd_present(*pgd))
291 goto out;
292 p4d = p4d_offset(pgd, address);
293 if (!p4d_present(*p4d))
294 goto out;
295 pud = pud_offset(p4d, address);
296 if (!pud_present(*pud))
297 goto out;
298 pmd = pmd_offset(pud, address);
299 /*
300 * READ_ONCE must function as a barrier with narrower scope
301 * and it must be equivalent to:
302 * _pmd = *pmd; barrier();
303 *
304 * This is to deal with the instability (as in
305 * pmd_trans_unstable) of the pmd.
306 */
307 _pmd = READ_ONCE(*pmd);
308 if (pmd_none(_pmd))
309 goto out;
310
311 ret = false;
312 if (!pmd_present(_pmd))
313 goto out;
314
315 if (pmd_trans_huge(_pmd)) {
316 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
317 ret = true;
318 goto out;
319 }
320
321 /*
322 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
323 * and use the standard pte_offset_map() instead of parsing _pmd.
324 */
325 pte = pte_offset_map(pmd, address);
326 /*
327 * Lockless access: we're in a wait_event so it's ok if it
328 * changes under us.
329 */
330 if (pte_none(*pte))
331 ret = true;
332 if (!pte_write(*pte) && (reason & VM_UFFD_WP))
333 ret = true;
334 pte_unmap(pte);
335
336out:
337 return ret;
338}
339
340static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
341{
342 if (flags & FAULT_FLAG_INTERRUPTIBLE)
343 return TASK_INTERRUPTIBLE;
344
345 if (flags & FAULT_FLAG_KILLABLE)
346 return TASK_KILLABLE;
347
348 return TASK_UNINTERRUPTIBLE;
349}
350
351/*
352 * The locking rules involved in returning VM_FAULT_RETRY depending on
353 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
354 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
355 * recommendation in __lock_page_or_retry is not an understatement.
356 *
357 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
358 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
359 * not set.
360 *
361 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
362 * set, VM_FAULT_RETRY can still be returned if and only if there are
363 * fatal_signal_pending()s, and the mmap_lock must be released before
364 * returning it.
365 */
366vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
367{
368 struct mm_struct *mm = vmf->vma->vm_mm;
369 struct userfaultfd_ctx *ctx;
370 struct userfaultfd_wait_queue uwq;
371 vm_fault_t ret = VM_FAULT_SIGBUS;
372 bool must_wait;
373 unsigned int blocking_state;
374
375 /*
376 * We don't do userfault handling for the final child pid update.
377 *
378 * We also don't do userfault handling during
379 * coredumping. hugetlbfs has the special
380 * follow_hugetlb_page() to skip missing pages in the
381 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
382 * the no_page_table() helper in follow_page_mask(), but the
383 * shmem_vm_ops->fault method is invoked even during
384 * coredumping without mmap_lock and it ends up here.
385 */
386 if (current->flags & (PF_EXITING|PF_DUMPCORE))
387 goto out;
388
389 /*
390 * Coredumping runs without mmap_lock so we can only check that
391 * the mmap_lock is held, if PF_DUMPCORE was not set.
392 */
393 mmap_assert_locked(mm);
394
395 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
396 if (!ctx)
397 goto out;
398
399 BUG_ON(ctx->mm != mm);
400
401 /* Any unrecognized flag is a bug. */
402 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
403 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
404 VM_BUG_ON(!reason || (reason & (reason - 1)));
405
406 if (ctx->features & UFFD_FEATURE_SIGBUS)
407 goto out;
408 if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
409 ctx->flags & UFFD_USER_MODE_ONLY) {
410 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
411 "sysctl knob to 1 if kernel faults must be handled "
412 "without obtaining CAP_SYS_PTRACE capability\n");
413 goto out;
414 }
415
416 /*
417 * If it's already released don't get it. This avoids to loop
418 * in __get_user_pages if userfaultfd_release waits on the
419 * caller of handle_userfault to release the mmap_lock.
420 */
421 if (unlikely(READ_ONCE(ctx->released))) {
422 /*
423 * Don't return VM_FAULT_SIGBUS in this case, so a non
424 * cooperative manager can close the uffd after the
425 * last UFFDIO_COPY, without risking to trigger an
426 * involuntary SIGBUS if the process was starting the
427 * userfaultfd while the userfaultfd was still armed
428 * (but after the last UFFDIO_COPY). If the uffd
429 * wasn't already closed when the userfault reached
430 * this point, that would normally be solved by
431 * userfaultfd_must_wait returning 'false'.
432 *
433 * If we were to return VM_FAULT_SIGBUS here, the non
434 * cooperative manager would be instead forced to
435 * always call UFFDIO_UNREGISTER before it can safely
436 * close the uffd.
437 */
438 ret = VM_FAULT_NOPAGE;
439 goto out;
440 }
441
442 /*
443 * Check that we can return VM_FAULT_RETRY.
444 *
445 * NOTE: it should become possible to return VM_FAULT_RETRY
446 * even if FAULT_FLAG_TRIED is set without leading to gup()
447 * -EBUSY failures, if the userfaultfd is to be extended for
448 * VM_UFFD_WP tracking and we intend to arm the userfault
449 * without first stopping userland access to the memory. For
450 * VM_UFFD_MISSING userfaults this is enough for now.
451 */
452 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
453 /*
454 * Validate the invariant that nowait must allow retry
455 * to be sure not to return SIGBUS erroneously on
456 * nowait invocations.
457 */
458 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
459#ifdef CONFIG_DEBUG_VM
460 if (printk_ratelimit()) {
461 printk(KERN_WARNING
462 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
463 vmf->flags);
464 dump_stack();
465 }
466#endif
467 goto out;
468 }
469
470 /*
471 * Handle nowait, not much to do other than tell it to retry
472 * and wait.
473 */
474 ret = VM_FAULT_RETRY;
475 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
476 goto out;
477
478 /* take the reference before dropping the mmap_lock */
479 userfaultfd_ctx_get(ctx);
480
481 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
482 uwq.wq.private = current;
483 uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
484 ctx->features);
485 uwq.ctx = ctx;
486 uwq.waken = false;
487
488 blocking_state = userfaultfd_get_blocking_state(vmf->flags);
489
490 spin_lock_irq(&ctx->fault_pending_wqh.lock);
491 /*
492 * After the __add_wait_queue the uwq is visible to userland
493 * through poll/read().
494 */
495 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
496 /*
497 * The smp_mb() after __set_current_state prevents the reads
498 * following the spin_unlock to happen before the list_add in
499 * __add_wait_queue.
500 */
501 set_current_state(blocking_state);
502 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
503
504 if (!is_vm_hugetlb_page(vmf->vma))
505 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
506 reason);
507 else
508 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
509 vmf->address,
510 vmf->flags, reason);
511 mmap_read_unlock(mm);
512
513 if (likely(must_wait && !READ_ONCE(ctx->released))) {
514 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
515 schedule();
516 }
517
518 __set_current_state(TASK_RUNNING);
519
520 /*
521 * Here we race with the list_del; list_add in
522 * userfaultfd_ctx_read(), however because we don't ever run
523 * list_del_init() to refile across the two lists, the prev
524 * and next pointers will never point to self. list_add also
525 * would never let any of the two pointers to point to
526 * self. So list_empty_careful won't risk to see both pointers
527 * pointing to self at any time during the list refile. The
528 * only case where list_del_init() is called is the full
529 * removal in the wake function and there we don't re-list_add
530 * and it's fine not to block on the spinlock. The uwq on this
531 * kernel stack can be released after the list_del_init.
532 */
533 if (!list_empty_careful(&uwq.wq.entry)) {
534 spin_lock_irq(&ctx->fault_pending_wqh.lock);
535 /*
536 * No need of list_del_init(), the uwq on the stack
537 * will be freed shortly anyway.
538 */
539 list_del(&uwq.wq.entry);
540 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
541 }
542
543 /*
544 * ctx may go away after this if the userfault pseudo fd is
545 * already released.
546 */
547 userfaultfd_ctx_put(ctx);
548
549out:
550 return ret;
551}
552
553static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
554 struct userfaultfd_wait_queue *ewq)
555{
556 struct userfaultfd_ctx *release_new_ctx;
557
558 if (WARN_ON_ONCE(current->flags & PF_EXITING))
559 goto out;
560
561 ewq->ctx = ctx;
562 init_waitqueue_entry(&ewq->wq, current);
563 release_new_ctx = NULL;
564
565 spin_lock_irq(&ctx->event_wqh.lock);
566 /*
567 * After the __add_wait_queue the uwq is visible to userland
568 * through poll/read().
569 */
570 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
571 for (;;) {
572 set_current_state(TASK_KILLABLE);
573 if (ewq->msg.event == 0)
574 break;
575 if (READ_ONCE(ctx->released) ||
576 fatal_signal_pending(current)) {
577 /*
578 * &ewq->wq may be queued in fork_event, but
579 * __remove_wait_queue ignores the head
580 * parameter. It would be a problem if it
581 * didn't.
582 */
583 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
584 if (ewq->msg.event == UFFD_EVENT_FORK) {
585 struct userfaultfd_ctx *new;
586
587 new = (struct userfaultfd_ctx *)
588 (unsigned long)
589 ewq->msg.arg.reserved.reserved1;
590 release_new_ctx = new;
591 }
592 break;
593 }
594
595 spin_unlock_irq(&ctx->event_wqh.lock);
596
597 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
598 schedule();
599
600 spin_lock_irq(&ctx->event_wqh.lock);
601 }
602 __set_current_state(TASK_RUNNING);
603 spin_unlock_irq(&ctx->event_wqh.lock);
604
605 if (release_new_ctx) {
606 struct vm_area_struct *vma;
607 struct mm_struct *mm = release_new_ctx->mm;
608
609 /* the various vma->vm_userfaultfd_ctx still points to it */
610 mmap_write_lock(mm);
611 for (vma = mm->mmap; vma; vma = vma->vm_next)
612 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
613 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
614 vma->vm_flags &= ~__VM_UFFD_FLAGS;
615 }
616 mmap_write_unlock(mm);
617
618 userfaultfd_ctx_put(release_new_ctx);
619 }
620
621 /*
622 * ctx may go away after this if the userfault pseudo fd is
623 * already released.
624 */
625out:
626 WRITE_ONCE(ctx->mmap_changing, false);
627 userfaultfd_ctx_put(ctx);
628}
629
630static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
631 struct userfaultfd_wait_queue *ewq)
632{
633 ewq->msg.event = 0;
634 wake_up_locked(&ctx->event_wqh);
635 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
636}
637
638int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
639{
640 struct userfaultfd_ctx *ctx = NULL, *octx;
641 struct userfaultfd_fork_ctx *fctx;
642
643 octx = vma->vm_userfaultfd_ctx.ctx;
644 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
645 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
646 vma->vm_flags &= ~__VM_UFFD_FLAGS;
647 return 0;
648 }
649
650 list_for_each_entry(fctx, fcs, list)
651 if (fctx->orig == octx) {
652 ctx = fctx->new;
653 break;
654 }
655
656 if (!ctx) {
657 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
658 if (!fctx)
659 return -ENOMEM;
660
661 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
662 if (!ctx) {
663 kfree(fctx);
664 return -ENOMEM;
665 }
666
667 refcount_set(&ctx->refcount, 1);
668 ctx->flags = octx->flags;
669 ctx->state = UFFD_STATE_RUNNING;
670 ctx->features = octx->features;
671 ctx->released = false;
672 ctx->mmap_changing = false;
673 ctx->mm = vma->vm_mm;
674 mmgrab(ctx->mm);
675
676 userfaultfd_ctx_get(octx);
677 WRITE_ONCE(octx->mmap_changing, true);
678 fctx->orig = octx;
679 fctx->new = ctx;
680 list_add_tail(&fctx->list, fcs);
681 }
682
683 vma->vm_userfaultfd_ctx.ctx = ctx;
684 return 0;
685}
686
687static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
688{
689 struct userfaultfd_ctx *ctx = fctx->orig;
690 struct userfaultfd_wait_queue ewq;
691
692 msg_init(&ewq.msg);
693
694 ewq.msg.event = UFFD_EVENT_FORK;
695 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
696
697 userfaultfd_event_wait_completion(ctx, &ewq);
698}
699
700void dup_userfaultfd_complete(struct list_head *fcs)
701{
702 struct userfaultfd_fork_ctx *fctx, *n;
703
704 list_for_each_entry_safe(fctx, n, fcs, list) {
705 dup_fctx(fctx);
706 list_del(&fctx->list);
707 kfree(fctx);
708 }
709}
710
711void mremap_userfaultfd_prep(struct vm_area_struct *vma,
712 struct vm_userfaultfd_ctx *vm_ctx)
713{
714 struct userfaultfd_ctx *ctx;
715
716 ctx = vma->vm_userfaultfd_ctx.ctx;
717
718 if (!ctx)
719 return;
720
721 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
722 vm_ctx->ctx = ctx;
723 userfaultfd_ctx_get(ctx);
724 WRITE_ONCE(ctx->mmap_changing, true);
725 } else {
726 /* Drop uffd context if remap feature not enabled */
727 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
728 vma->vm_flags &= ~__VM_UFFD_FLAGS;
729 }
730}
731
732void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
733 unsigned long from, unsigned long to,
734 unsigned long len)
735{
736 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
737 struct userfaultfd_wait_queue ewq;
738
739 if (!ctx)
740 return;
741
742 if (to & ~PAGE_MASK) {
743 userfaultfd_ctx_put(ctx);
744 return;
745 }
746
747 msg_init(&ewq.msg);
748
749 ewq.msg.event = UFFD_EVENT_REMAP;
750 ewq.msg.arg.remap.from = from;
751 ewq.msg.arg.remap.to = to;
752 ewq.msg.arg.remap.len = len;
753
754 userfaultfd_event_wait_completion(ctx, &ewq);
755}
756
757bool userfaultfd_remove(struct vm_area_struct *vma,
758 unsigned long start, unsigned long end)
759{
760 struct mm_struct *mm = vma->vm_mm;
761 struct userfaultfd_ctx *ctx;
762 struct userfaultfd_wait_queue ewq;
763
764 ctx = vma->vm_userfaultfd_ctx.ctx;
765 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
766 return true;
767
768 userfaultfd_ctx_get(ctx);
769 WRITE_ONCE(ctx->mmap_changing, true);
770 mmap_read_unlock(mm);
771
772 msg_init(&ewq.msg);
773
774 ewq.msg.event = UFFD_EVENT_REMOVE;
775 ewq.msg.arg.remove.start = start;
776 ewq.msg.arg.remove.end = end;
777
778 userfaultfd_event_wait_completion(ctx, &ewq);
779
780 return false;
781}
782
783static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
784 unsigned long start, unsigned long end)
785{
786 struct userfaultfd_unmap_ctx *unmap_ctx;
787
788 list_for_each_entry(unmap_ctx, unmaps, list)
789 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
790 unmap_ctx->end == end)
791 return true;
792
793 return false;
794}
795
796int userfaultfd_unmap_prep(struct vm_area_struct *vma,
797 unsigned long start, unsigned long end,
798 struct list_head *unmaps)
799{
800 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
801 struct userfaultfd_unmap_ctx *unmap_ctx;
802 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
803
804 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
805 has_unmap_ctx(ctx, unmaps, start, end))
806 continue;
807
808 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
809 if (!unmap_ctx)
810 return -ENOMEM;
811
812 userfaultfd_ctx_get(ctx);
813 WRITE_ONCE(ctx->mmap_changing, true);
814 unmap_ctx->ctx = ctx;
815 unmap_ctx->start = start;
816 unmap_ctx->end = end;
817 list_add_tail(&unmap_ctx->list, unmaps);
818 }
819
820 return 0;
821}
822
823void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
824{
825 struct userfaultfd_unmap_ctx *ctx, *n;
826 struct userfaultfd_wait_queue ewq;
827
828 list_for_each_entry_safe(ctx, n, uf, list) {
829 msg_init(&ewq.msg);
830
831 ewq.msg.event = UFFD_EVENT_UNMAP;
832 ewq.msg.arg.remove.start = ctx->start;
833 ewq.msg.arg.remove.end = ctx->end;
834
835 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
836
837 list_del(&ctx->list);
838 kfree(ctx);
839 }
840}
841
842static int userfaultfd_release(struct inode *inode, struct file *file)
843{
844 struct userfaultfd_ctx *ctx = file->private_data;
845 struct mm_struct *mm = ctx->mm;
846 struct vm_area_struct *vma, *prev;
847 /* len == 0 means wake all */
848 struct userfaultfd_wake_range range = { .len = 0, };
849 unsigned long new_flags;
850
851 WRITE_ONCE(ctx->released, true);
852
853 if (!mmget_not_zero(mm))
854 goto wakeup;
855
856 /*
857 * Flush page faults out of all CPUs. NOTE: all page faults
858 * must be retried without returning VM_FAULT_SIGBUS if
859 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
860 * changes while handle_userfault released the mmap_lock. So
861 * it's critical that released is set to true (above), before
862 * taking the mmap_lock for writing.
863 */
864 mmap_write_lock(mm);
865 prev = NULL;
866 for (vma = mm->mmap; vma; vma = vma->vm_next) {
867 cond_resched();
868 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
869 !!(vma->vm_flags & __VM_UFFD_FLAGS));
870 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
871 prev = vma;
872 continue;
873 }
874 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
875 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
876 new_flags, vma->anon_vma,
877 vma->vm_file, vma->vm_pgoff,
878 vma_policy(vma),
879 NULL_VM_UFFD_CTX);
880 if (prev)
881 vma = prev;
882 else
883 prev = vma;
884 vma->vm_flags = new_flags;
885 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
886 }
887 mmap_write_unlock(mm);
888 mmput(mm);
889wakeup:
890 /*
891 * After no new page faults can wait on this fault_*wqh, flush
892 * the last page faults that may have been already waiting on
893 * the fault_*wqh.
894 */
895 spin_lock_irq(&ctx->fault_pending_wqh.lock);
896 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
897 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
898 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
899
900 /* Flush pending events that may still wait on event_wqh */
901 wake_up_all(&ctx->event_wqh);
902
903 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
904 userfaultfd_ctx_put(ctx);
905 return 0;
906}
907
908/* fault_pending_wqh.lock must be hold by the caller */
909static inline struct userfaultfd_wait_queue *find_userfault_in(
910 wait_queue_head_t *wqh)
911{
912 wait_queue_entry_t *wq;
913 struct userfaultfd_wait_queue *uwq;
914
915 lockdep_assert_held(&wqh->lock);
916
917 uwq = NULL;
918 if (!waitqueue_active(wqh))
919 goto out;
920 /* walk in reverse to provide FIFO behavior to read userfaults */
921 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
922 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
923out:
924 return uwq;
925}
926
927static inline struct userfaultfd_wait_queue *find_userfault(
928 struct userfaultfd_ctx *ctx)
929{
930 return find_userfault_in(&ctx->fault_pending_wqh);
931}
932
933static inline struct userfaultfd_wait_queue *find_userfault_evt(
934 struct userfaultfd_ctx *ctx)
935{
936 return find_userfault_in(&ctx->event_wqh);
937}
938
939static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
940{
941 struct userfaultfd_ctx *ctx = file->private_data;
942 __poll_t ret;
943
944 poll_wait(file, &ctx->fd_wqh, wait);
945
946 switch (ctx->state) {
947 case UFFD_STATE_WAIT_API:
948 return EPOLLERR;
949 case UFFD_STATE_RUNNING:
950 /*
951 * poll() never guarantees that read won't block.
952 * userfaults can be waken before they're read().
953 */
954 if (unlikely(!(file->f_flags & O_NONBLOCK)))
955 return EPOLLERR;
956 /*
957 * lockless access to see if there are pending faults
958 * __pollwait last action is the add_wait_queue but
959 * the spin_unlock would allow the waitqueue_active to
960 * pass above the actual list_add inside
961 * add_wait_queue critical section. So use a full
962 * memory barrier to serialize the list_add write of
963 * add_wait_queue() with the waitqueue_active read
964 * below.
965 */
966 ret = 0;
967 smp_mb();
968 if (waitqueue_active(&ctx->fault_pending_wqh))
969 ret = EPOLLIN;
970 else if (waitqueue_active(&ctx->event_wqh))
971 ret = EPOLLIN;
972
973 return ret;
974 default:
975 WARN_ON_ONCE(1);
976 return EPOLLERR;
977 }
978}
979
980static const struct file_operations userfaultfd_fops;
981
982static int resolve_userfault_fork(struct userfaultfd_ctx *new,
983 struct inode *inode,
984 struct uffd_msg *msg)
985{
986 int fd;
987
988 fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
989 O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
990 if (fd < 0)
991 return fd;
992
993 msg->arg.reserved.reserved1 = 0;
994 msg->arg.fork.ufd = fd;
995 return 0;
996}
997
998static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
999 struct uffd_msg *msg, struct inode *inode)
1000{
1001 ssize_t ret;
1002 DECLARE_WAITQUEUE(wait, current);
1003 struct userfaultfd_wait_queue *uwq;
1004 /*
1005 * Handling fork event requires sleeping operations, so
1006 * we drop the event_wqh lock, then do these ops, then
1007 * lock it back and wake up the waiter. While the lock is
1008 * dropped the ewq may go away so we keep track of it
1009 * carefully.
1010 */
1011 LIST_HEAD(fork_event);
1012 struct userfaultfd_ctx *fork_nctx = NULL;
1013
1014 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1015 spin_lock_irq(&ctx->fd_wqh.lock);
1016 __add_wait_queue(&ctx->fd_wqh, &wait);
1017 for (;;) {
1018 set_current_state(TASK_INTERRUPTIBLE);
1019 spin_lock(&ctx->fault_pending_wqh.lock);
1020 uwq = find_userfault(ctx);
1021 if (uwq) {
1022 /*
1023 * Use a seqcount to repeat the lockless check
1024 * in wake_userfault() to avoid missing
1025 * wakeups because during the refile both
1026 * waitqueue could become empty if this is the
1027 * only userfault.
1028 */
1029 write_seqcount_begin(&ctx->refile_seq);
1030
1031 /*
1032 * The fault_pending_wqh.lock prevents the uwq
1033 * to disappear from under us.
1034 *
1035 * Refile this userfault from
1036 * fault_pending_wqh to fault_wqh, it's not
1037 * pending anymore after we read it.
1038 *
1039 * Use list_del() by hand (as
1040 * userfaultfd_wake_function also uses
1041 * list_del_init() by hand) to be sure nobody
1042 * changes __remove_wait_queue() to use
1043 * list_del_init() in turn breaking the
1044 * !list_empty_careful() check in
1045 * handle_userfault(). The uwq->wq.head list
1046 * must never be empty at any time during the
1047 * refile, or the waitqueue could disappear
1048 * from under us. The "wait_queue_head_t"
1049 * parameter of __remove_wait_queue() is unused
1050 * anyway.
1051 */
1052 list_del(&uwq->wq.entry);
1053 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1054
1055 write_seqcount_end(&ctx->refile_seq);
1056
1057 /* careful to always initialize msg if ret == 0 */
1058 *msg = uwq->msg;
1059 spin_unlock(&ctx->fault_pending_wqh.lock);
1060 ret = 0;
1061 break;
1062 }
1063 spin_unlock(&ctx->fault_pending_wqh.lock);
1064
1065 spin_lock(&ctx->event_wqh.lock);
1066 uwq = find_userfault_evt(ctx);
1067 if (uwq) {
1068 *msg = uwq->msg;
1069
1070 if (uwq->msg.event == UFFD_EVENT_FORK) {
1071 fork_nctx = (struct userfaultfd_ctx *)
1072 (unsigned long)
1073 uwq->msg.arg.reserved.reserved1;
1074 list_move(&uwq->wq.entry, &fork_event);
1075 /*
1076 * fork_nctx can be freed as soon as
1077 * we drop the lock, unless we take a
1078 * reference on it.
1079 */
1080 userfaultfd_ctx_get(fork_nctx);
1081 spin_unlock(&ctx->event_wqh.lock);
1082 ret = 0;
1083 break;
1084 }
1085
1086 userfaultfd_event_complete(ctx, uwq);
1087 spin_unlock(&ctx->event_wqh.lock);
1088 ret = 0;
1089 break;
1090 }
1091 spin_unlock(&ctx->event_wqh.lock);
1092
1093 if (signal_pending(current)) {
1094 ret = -ERESTARTSYS;
1095 break;
1096 }
1097 if (no_wait) {
1098 ret = -EAGAIN;
1099 break;
1100 }
1101 spin_unlock_irq(&ctx->fd_wqh.lock);
1102 schedule();
1103 spin_lock_irq(&ctx->fd_wqh.lock);
1104 }
1105 __remove_wait_queue(&ctx->fd_wqh, &wait);
1106 __set_current_state(TASK_RUNNING);
1107 spin_unlock_irq(&ctx->fd_wqh.lock);
1108
1109 if (!ret && msg->event == UFFD_EVENT_FORK) {
1110 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1111 spin_lock_irq(&ctx->event_wqh.lock);
1112 if (!list_empty(&fork_event)) {
1113 /*
1114 * The fork thread didn't abort, so we can
1115 * drop the temporary refcount.
1116 */
1117 userfaultfd_ctx_put(fork_nctx);
1118
1119 uwq = list_first_entry(&fork_event,
1120 typeof(*uwq),
1121 wq.entry);
1122 /*
1123 * If fork_event list wasn't empty and in turn
1124 * the event wasn't already released by fork
1125 * (the event is allocated on fork kernel
1126 * stack), put the event back to its place in
1127 * the event_wq. fork_event head will be freed
1128 * as soon as we return so the event cannot
1129 * stay queued there no matter the current
1130 * "ret" value.
1131 */
1132 list_del(&uwq->wq.entry);
1133 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1134
1135 /*
1136 * Leave the event in the waitqueue and report
1137 * error to userland if we failed to resolve
1138 * the userfault fork.
1139 */
1140 if (likely(!ret))
1141 userfaultfd_event_complete(ctx, uwq);
1142 } else {
1143 /*
1144 * Here the fork thread aborted and the
1145 * refcount from the fork thread on fork_nctx
1146 * has already been released. We still hold
1147 * the reference we took before releasing the
1148 * lock above. If resolve_userfault_fork
1149 * failed we've to drop it because the
1150 * fork_nctx has to be freed in such case. If
1151 * it succeeded we'll hold it because the new
1152 * uffd references it.
1153 */
1154 if (ret)
1155 userfaultfd_ctx_put(fork_nctx);
1156 }
1157 spin_unlock_irq(&ctx->event_wqh.lock);
1158 }
1159
1160 return ret;
1161}
1162
1163static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1164 size_t count, loff_t *ppos)
1165{
1166 struct userfaultfd_ctx *ctx = file->private_data;
1167 ssize_t _ret, ret = 0;
1168 struct uffd_msg msg;
1169 int no_wait = file->f_flags & O_NONBLOCK;
1170 struct inode *inode = file_inode(file);
1171
1172 if (ctx->state == UFFD_STATE_WAIT_API)
1173 return -EINVAL;
1174
1175 for (;;) {
1176 if (count < sizeof(msg))
1177 return ret ? ret : -EINVAL;
1178 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1179 if (_ret < 0)
1180 return ret ? ret : _ret;
1181 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1182 return ret ? ret : -EFAULT;
1183 ret += sizeof(msg);
1184 buf += sizeof(msg);
1185 count -= sizeof(msg);
1186 /*
1187 * Allow to read more than one fault at time but only
1188 * block if waiting for the very first one.
1189 */
1190 no_wait = O_NONBLOCK;
1191 }
1192}
1193
1194static void __wake_userfault(struct userfaultfd_ctx *ctx,
1195 struct userfaultfd_wake_range *range)
1196{
1197 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1198 /* wake all in the range and autoremove */
1199 if (waitqueue_active(&ctx->fault_pending_wqh))
1200 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1201 range);
1202 if (waitqueue_active(&ctx->fault_wqh))
1203 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1204 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1205}
1206
1207static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1208 struct userfaultfd_wake_range *range)
1209{
1210 unsigned seq;
1211 bool need_wakeup;
1212
1213 /*
1214 * To be sure waitqueue_active() is not reordered by the CPU
1215 * before the pagetable update, use an explicit SMP memory
1216 * barrier here. PT lock release or mmap_read_unlock(mm) still
1217 * have release semantics that can allow the
1218 * waitqueue_active() to be reordered before the pte update.
1219 */
1220 smp_mb();
1221
1222 /*
1223 * Use waitqueue_active because it's very frequent to
1224 * change the address space atomically even if there are no
1225 * userfaults yet. So we take the spinlock only when we're
1226 * sure we've userfaults to wake.
1227 */
1228 do {
1229 seq = read_seqcount_begin(&ctx->refile_seq);
1230 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1231 waitqueue_active(&ctx->fault_wqh);
1232 cond_resched();
1233 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1234 if (need_wakeup)
1235 __wake_userfault(ctx, range);
1236}
1237
1238static __always_inline int validate_range(struct mm_struct *mm,
1239 __u64 start, __u64 len)
1240{
1241 __u64 task_size = mm->task_size;
1242
1243 if (start & ~PAGE_MASK)
1244 return -EINVAL;
1245 if (len & ~PAGE_MASK)
1246 return -EINVAL;
1247 if (!len)
1248 return -EINVAL;
1249 if (start < mmap_min_addr)
1250 return -EINVAL;
1251 if (start >= task_size)
1252 return -EINVAL;
1253 if (len > task_size - start)
1254 return -EINVAL;
1255 return 0;
1256}
1257
1258static inline bool vma_can_userfault(struct vm_area_struct *vma,
1259 unsigned long vm_flags)
1260{
1261 /* FIXME: add WP support to hugetlbfs and shmem */
1262 if (vm_flags & VM_UFFD_WP) {
1263 if (is_vm_hugetlb_page(vma) || vma_is_shmem(vma))
1264 return false;
1265 }
1266
1267 if (vm_flags & VM_UFFD_MINOR) {
1268 if (!(is_vm_hugetlb_page(vma) || vma_is_shmem(vma)))
1269 return false;
1270 }
1271
1272 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1273 vma_is_shmem(vma);
1274}
1275
1276static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1277 unsigned long arg)
1278{
1279 struct mm_struct *mm = ctx->mm;
1280 struct vm_area_struct *vma, *prev, *cur;
1281 int ret;
1282 struct uffdio_register uffdio_register;
1283 struct uffdio_register __user *user_uffdio_register;
1284 unsigned long vm_flags, new_flags;
1285 bool found;
1286 bool basic_ioctls;
1287 unsigned long start, end, vma_end;
1288
1289 user_uffdio_register = (struct uffdio_register __user *) arg;
1290
1291 ret = -EFAULT;
1292 if (copy_from_user(&uffdio_register, user_uffdio_register,
1293 sizeof(uffdio_register)-sizeof(__u64)))
1294 goto out;
1295
1296 ret = -EINVAL;
1297 if (!uffdio_register.mode)
1298 goto out;
1299 if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1300 goto out;
1301 vm_flags = 0;
1302 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1303 vm_flags |= VM_UFFD_MISSING;
1304 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1305#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1306 goto out;
1307#endif
1308 vm_flags |= VM_UFFD_WP;
1309 }
1310 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1311#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1312 goto out;
1313#endif
1314 vm_flags |= VM_UFFD_MINOR;
1315 }
1316
1317 ret = validate_range(mm, uffdio_register.range.start,
1318 uffdio_register.range.len);
1319 if (ret)
1320 goto out;
1321
1322 start = uffdio_register.range.start;
1323 end = start + uffdio_register.range.len;
1324
1325 ret = -ENOMEM;
1326 if (!mmget_not_zero(mm))
1327 goto out;
1328
1329 mmap_write_lock(mm);
1330 vma = find_vma_prev(mm, start, &prev);
1331 if (!vma)
1332 goto out_unlock;
1333
1334 /* check that there's at least one vma in the range */
1335 ret = -EINVAL;
1336 if (vma->vm_start >= end)
1337 goto out_unlock;
1338
1339 /*
1340 * If the first vma contains huge pages, make sure start address
1341 * is aligned to huge page size.
1342 */
1343 if (is_vm_hugetlb_page(vma)) {
1344 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1345
1346 if (start & (vma_hpagesize - 1))
1347 goto out_unlock;
1348 }
1349
1350 /*
1351 * Search for not compatible vmas.
1352 */
1353 found = false;
1354 basic_ioctls = false;
1355 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1356 cond_resched();
1357
1358 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1359 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1360
1361 /* check not compatible vmas */
1362 ret = -EINVAL;
1363 if (!vma_can_userfault(cur, vm_flags))
1364 goto out_unlock;
1365
1366 /*
1367 * UFFDIO_COPY will fill file holes even without
1368 * PROT_WRITE. This check enforces that if this is a
1369 * MAP_SHARED, the process has write permission to the backing
1370 * file. If VM_MAYWRITE is set it also enforces that on a
1371 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1372 * F_WRITE_SEAL can be taken until the vma is destroyed.
1373 */
1374 ret = -EPERM;
1375 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1376 goto out_unlock;
1377
1378 /*
1379 * If this vma contains ending address, and huge pages
1380 * check alignment.
1381 */
1382 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1383 end > cur->vm_start) {
1384 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1385
1386 ret = -EINVAL;
1387
1388 if (end & (vma_hpagesize - 1))
1389 goto out_unlock;
1390 }
1391 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1392 goto out_unlock;
1393
1394 /*
1395 * Check that this vma isn't already owned by a
1396 * different userfaultfd. We can't allow more than one
1397 * userfaultfd to own a single vma simultaneously or we
1398 * wouldn't know which one to deliver the userfaults to.
1399 */
1400 ret = -EBUSY;
1401 if (cur->vm_userfaultfd_ctx.ctx &&
1402 cur->vm_userfaultfd_ctx.ctx != ctx)
1403 goto out_unlock;
1404
1405 /*
1406 * Note vmas containing huge pages
1407 */
1408 if (is_vm_hugetlb_page(cur))
1409 basic_ioctls = true;
1410
1411 found = true;
1412 }
1413 BUG_ON(!found);
1414
1415 if (vma->vm_start < start)
1416 prev = vma;
1417
1418 ret = 0;
1419 do {
1420 cond_resched();
1421
1422 BUG_ON(!vma_can_userfault(vma, vm_flags));
1423 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1424 vma->vm_userfaultfd_ctx.ctx != ctx);
1425 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1426
1427 /*
1428 * Nothing to do: this vma is already registered into this
1429 * userfaultfd and with the right tracking mode too.
1430 */
1431 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1432 (vma->vm_flags & vm_flags) == vm_flags)
1433 goto skip;
1434
1435 if (vma->vm_start > start)
1436 start = vma->vm_start;
1437 vma_end = min(end, vma->vm_end);
1438
1439 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1440 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1441 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1442 vma_policy(vma),
1443 ((struct vm_userfaultfd_ctx){ ctx }));
1444 if (prev) {
1445 vma = prev;
1446 goto next;
1447 }
1448 if (vma->vm_start < start) {
1449 ret = split_vma(mm, vma, start, 1);
1450 if (ret)
1451 break;
1452 }
1453 if (vma->vm_end > end) {
1454 ret = split_vma(mm, vma, end, 0);
1455 if (ret)
1456 break;
1457 }
1458 next:
1459 /*
1460 * In the vma_merge() successful mprotect-like case 8:
1461 * the next vma was merged into the current one and
1462 * the current one has not been updated yet.
1463 */
1464 vma->vm_flags = new_flags;
1465 vma->vm_userfaultfd_ctx.ctx = ctx;
1466
1467 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1468 hugetlb_unshare_all_pmds(vma);
1469
1470 skip:
1471 prev = vma;
1472 start = vma->vm_end;
1473 vma = vma->vm_next;
1474 } while (vma && vma->vm_start < end);
1475out_unlock:
1476 mmap_write_unlock(mm);
1477 mmput(mm);
1478 if (!ret) {
1479 __u64 ioctls_out;
1480
1481 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1482 UFFD_API_RANGE_IOCTLS;
1483
1484 /*
1485 * Declare the WP ioctl only if the WP mode is
1486 * specified and all checks passed with the range
1487 */
1488 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1489 ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1490
1491 /* CONTINUE ioctl is only supported for MINOR ranges. */
1492 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1493 ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1494
1495 /*
1496 * Now that we scanned all vmas we can already tell
1497 * userland which ioctls methods are guaranteed to
1498 * succeed on this range.
1499 */
1500 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1501 ret = -EFAULT;
1502 }
1503out:
1504 return ret;
1505}
1506
1507static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1508 unsigned long arg)
1509{
1510 struct mm_struct *mm = ctx->mm;
1511 struct vm_area_struct *vma, *prev, *cur;
1512 int ret;
1513 struct uffdio_range uffdio_unregister;
1514 unsigned long new_flags;
1515 bool found;
1516 unsigned long start, end, vma_end;
1517 const void __user *buf = (void __user *)arg;
1518
1519 ret = -EFAULT;
1520 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1521 goto out;
1522
1523 ret = validate_range(mm, uffdio_unregister.start,
1524 uffdio_unregister.len);
1525 if (ret)
1526 goto out;
1527
1528 start = uffdio_unregister.start;
1529 end = start + uffdio_unregister.len;
1530
1531 ret = -ENOMEM;
1532 if (!mmget_not_zero(mm))
1533 goto out;
1534
1535 mmap_write_lock(mm);
1536 vma = find_vma_prev(mm, start, &prev);
1537 if (!vma)
1538 goto out_unlock;
1539
1540 /* check that there's at least one vma in the range */
1541 ret = -EINVAL;
1542 if (vma->vm_start >= end)
1543 goto out_unlock;
1544
1545 /*
1546 * If the first vma contains huge pages, make sure start address
1547 * is aligned to huge page size.
1548 */
1549 if (is_vm_hugetlb_page(vma)) {
1550 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1551
1552 if (start & (vma_hpagesize - 1))
1553 goto out_unlock;
1554 }
1555
1556 /*
1557 * Search for not compatible vmas.
1558 */
1559 found = false;
1560 ret = -EINVAL;
1561 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1562 cond_resched();
1563
1564 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1565 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1566
1567 /*
1568 * Check not compatible vmas, not strictly required
1569 * here as not compatible vmas cannot have an
1570 * userfaultfd_ctx registered on them, but this
1571 * provides for more strict behavior to notice
1572 * unregistration errors.
1573 */
1574 if (!vma_can_userfault(cur, cur->vm_flags))
1575 goto out_unlock;
1576
1577 found = true;
1578 }
1579 BUG_ON(!found);
1580
1581 if (vma->vm_start < start)
1582 prev = vma;
1583
1584 ret = 0;
1585 do {
1586 cond_resched();
1587
1588 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1589
1590 /*
1591 * Nothing to do: this vma is already registered into this
1592 * userfaultfd and with the right tracking mode too.
1593 */
1594 if (!vma->vm_userfaultfd_ctx.ctx)
1595 goto skip;
1596
1597 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1598
1599 if (vma->vm_start > start)
1600 start = vma->vm_start;
1601 vma_end = min(end, vma->vm_end);
1602
1603 if (userfaultfd_missing(vma)) {
1604 /*
1605 * Wake any concurrent pending userfault while
1606 * we unregister, so they will not hang
1607 * permanently and it avoids userland to call
1608 * UFFDIO_WAKE explicitly.
1609 */
1610 struct userfaultfd_wake_range range;
1611 range.start = start;
1612 range.len = vma_end - start;
1613 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1614 }
1615
1616 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1617 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1618 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1619 vma_policy(vma),
1620 NULL_VM_UFFD_CTX);
1621 if (prev) {
1622 vma = prev;
1623 goto next;
1624 }
1625 if (vma->vm_start < start) {
1626 ret = split_vma(mm, vma, start, 1);
1627 if (ret)
1628 break;
1629 }
1630 if (vma->vm_end > end) {
1631 ret = split_vma(mm, vma, end, 0);
1632 if (ret)
1633 break;
1634 }
1635 next:
1636 /*
1637 * In the vma_merge() successful mprotect-like case 8:
1638 * the next vma was merged into the current one and
1639 * the current one has not been updated yet.
1640 */
1641 vma->vm_flags = new_flags;
1642 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1643
1644 skip:
1645 prev = vma;
1646 start = vma->vm_end;
1647 vma = vma->vm_next;
1648 } while (vma && vma->vm_start < end);
1649out_unlock:
1650 mmap_write_unlock(mm);
1651 mmput(mm);
1652out:
1653 return ret;
1654}
1655
1656/*
1657 * userfaultfd_wake may be used in combination with the
1658 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1659 */
1660static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1661 unsigned long arg)
1662{
1663 int ret;
1664 struct uffdio_range uffdio_wake;
1665 struct userfaultfd_wake_range range;
1666 const void __user *buf = (void __user *)arg;
1667
1668 ret = -EFAULT;
1669 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1670 goto out;
1671
1672 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1673 if (ret)
1674 goto out;
1675
1676 range.start = uffdio_wake.start;
1677 range.len = uffdio_wake.len;
1678
1679 /*
1680 * len == 0 means wake all and we don't want to wake all here,
1681 * so check it again to be sure.
1682 */
1683 VM_BUG_ON(!range.len);
1684
1685 wake_userfault(ctx, &range);
1686 ret = 0;
1687
1688out:
1689 return ret;
1690}
1691
1692static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1693 unsigned long arg)
1694{
1695 __s64 ret;
1696 struct uffdio_copy uffdio_copy;
1697 struct uffdio_copy __user *user_uffdio_copy;
1698 struct userfaultfd_wake_range range;
1699
1700 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1701
1702 ret = -EAGAIN;
1703 if (READ_ONCE(ctx->mmap_changing))
1704 goto out;
1705
1706 ret = -EFAULT;
1707 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1708 /* don't copy "copy" last field */
1709 sizeof(uffdio_copy)-sizeof(__s64)))
1710 goto out;
1711
1712 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1713 if (ret)
1714 goto out;
1715 /*
1716 * double check for wraparound just in case. copy_from_user()
1717 * will later check uffdio_copy.src + uffdio_copy.len to fit
1718 * in the userland range.
1719 */
1720 ret = -EINVAL;
1721 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1722 goto out;
1723 if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1724 goto out;
1725 if (mmget_not_zero(ctx->mm)) {
1726 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1727 uffdio_copy.len, &ctx->mmap_changing,
1728 uffdio_copy.mode);
1729 mmput(ctx->mm);
1730 } else {
1731 return -ESRCH;
1732 }
1733 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1734 return -EFAULT;
1735 if (ret < 0)
1736 goto out;
1737 BUG_ON(!ret);
1738 /* len == 0 would wake all */
1739 range.len = ret;
1740 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1741 range.start = uffdio_copy.dst;
1742 wake_userfault(ctx, &range);
1743 }
1744 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1745out:
1746 return ret;
1747}
1748
1749static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1750 unsigned long arg)
1751{
1752 __s64 ret;
1753 struct uffdio_zeropage uffdio_zeropage;
1754 struct uffdio_zeropage __user *user_uffdio_zeropage;
1755 struct userfaultfd_wake_range range;
1756
1757 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1758
1759 ret = -EAGAIN;
1760 if (READ_ONCE(ctx->mmap_changing))
1761 goto out;
1762
1763 ret = -EFAULT;
1764 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1765 /* don't copy "zeropage" last field */
1766 sizeof(uffdio_zeropage)-sizeof(__s64)))
1767 goto out;
1768
1769 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1770 uffdio_zeropage.range.len);
1771 if (ret)
1772 goto out;
1773 ret = -EINVAL;
1774 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1775 goto out;
1776
1777 if (mmget_not_zero(ctx->mm)) {
1778 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1779 uffdio_zeropage.range.len,
1780 &ctx->mmap_changing);
1781 mmput(ctx->mm);
1782 } else {
1783 return -ESRCH;
1784 }
1785 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1786 return -EFAULT;
1787 if (ret < 0)
1788 goto out;
1789 /* len == 0 would wake all */
1790 BUG_ON(!ret);
1791 range.len = ret;
1792 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1793 range.start = uffdio_zeropage.range.start;
1794 wake_userfault(ctx, &range);
1795 }
1796 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1797out:
1798 return ret;
1799}
1800
1801static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1802 unsigned long arg)
1803{
1804 int ret;
1805 struct uffdio_writeprotect uffdio_wp;
1806 struct uffdio_writeprotect __user *user_uffdio_wp;
1807 struct userfaultfd_wake_range range;
1808 bool mode_wp, mode_dontwake;
1809
1810 if (READ_ONCE(ctx->mmap_changing))
1811 return -EAGAIN;
1812
1813 user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1814
1815 if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1816 sizeof(struct uffdio_writeprotect)))
1817 return -EFAULT;
1818
1819 ret = validate_range(ctx->mm, uffdio_wp.range.start,
1820 uffdio_wp.range.len);
1821 if (ret)
1822 return ret;
1823
1824 if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1825 UFFDIO_WRITEPROTECT_MODE_WP))
1826 return -EINVAL;
1827
1828 mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1829 mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1830
1831 if (mode_wp && mode_dontwake)
1832 return -EINVAL;
1833
1834 ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1835 uffdio_wp.range.len, mode_wp,
1836 &ctx->mmap_changing);
1837 if (ret)
1838 return ret;
1839
1840 if (!mode_wp && !mode_dontwake) {
1841 range.start = uffdio_wp.range.start;
1842 range.len = uffdio_wp.range.len;
1843 wake_userfault(ctx, &range);
1844 }
1845 return ret;
1846}
1847
1848static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1849{
1850 __s64 ret;
1851 struct uffdio_continue uffdio_continue;
1852 struct uffdio_continue __user *user_uffdio_continue;
1853 struct userfaultfd_wake_range range;
1854
1855 user_uffdio_continue = (struct uffdio_continue __user *)arg;
1856
1857 ret = -EAGAIN;
1858 if (READ_ONCE(ctx->mmap_changing))
1859 goto out;
1860
1861 ret = -EFAULT;
1862 if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1863 /* don't copy the output fields */
1864 sizeof(uffdio_continue) - (sizeof(__s64))))
1865 goto out;
1866
1867 ret = validate_range(ctx->mm, uffdio_continue.range.start,
1868 uffdio_continue.range.len);
1869 if (ret)
1870 goto out;
1871
1872 ret = -EINVAL;
1873 /* double check for wraparound just in case. */
1874 if (uffdio_continue.range.start + uffdio_continue.range.len <=
1875 uffdio_continue.range.start) {
1876 goto out;
1877 }
1878 if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1879 goto out;
1880
1881 if (mmget_not_zero(ctx->mm)) {
1882 ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1883 uffdio_continue.range.len,
1884 &ctx->mmap_changing);
1885 mmput(ctx->mm);
1886 } else {
1887 return -ESRCH;
1888 }
1889
1890 if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1891 return -EFAULT;
1892 if (ret < 0)
1893 goto out;
1894
1895 /* len == 0 would wake all */
1896 BUG_ON(!ret);
1897 range.len = ret;
1898 if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1899 range.start = uffdio_continue.range.start;
1900 wake_userfault(ctx, &range);
1901 }
1902 ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1903
1904out:
1905 return ret;
1906}
1907
1908static inline unsigned int uffd_ctx_features(__u64 user_features)
1909{
1910 /*
1911 * For the current set of features the bits just coincide
1912 */
1913 return (unsigned int)user_features;
1914}
1915
1916/*
1917 * userland asks for a certain API version and we return which bits
1918 * and ioctl commands are implemented in this kernel for such API
1919 * version or -EINVAL if unknown.
1920 */
1921static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1922 unsigned long arg)
1923{
1924 struct uffdio_api uffdio_api;
1925 void __user *buf = (void __user *)arg;
1926 int ret;
1927 __u64 features;
1928
1929 ret = -EINVAL;
1930 if (ctx->state != UFFD_STATE_WAIT_API)
1931 goto out;
1932 ret = -EFAULT;
1933 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1934 goto out;
1935 features = uffdio_api.features;
1936 ret = -EINVAL;
1937 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1938 goto err_out;
1939 ret = -EPERM;
1940 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1941 goto err_out;
1942 /* report all available features and ioctls to userland */
1943 uffdio_api.features = UFFD_API_FEATURES;
1944#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1945 uffdio_api.features &=
1946 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1947#endif
1948#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1949 uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1950#endif
1951 uffdio_api.ioctls = UFFD_API_IOCTLS;
1952 ret = -EFAULT;
1953 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1954 goto out;
1955 ctx->state = UFFD_STATE_RUNNING;
1956 /* only enable the requested features for this uffd context */
1957 ctx->features = uffd_ctx_features(features);
1958 ret = 0;
1959out:
1960 return ret;
1961err_out:
1962 memset(&uffdio_api, 0, sizeof(uffdio_api));
1963 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1964 ret = -EFAULT;
1965 goto out;
1966}
1967
1968static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1969 unsigned long arg)
1970{
1971 int ret = -EINVAL;
1972 struct userfaultfd_ctx *ctx = file->private_data;
1973
1974 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1975 return -EINVAL;
1976
1977 switch(cmd) {
1978 case UFFDIO_API:
1979 ret = userfaultfd_api(ctx, arg);
1980 break;
1981 case UFFDIO_REGISTER:
1982 ret = userfaultfd_register(ctx, arg);
1983 break;
1984 case UFFDIO_UNREGISTER:
1985 ret = userfaultfd_unregister(ctx, arg);
1986 break;
1987 case UFFDIO_WAKE:
1988 ret = userfaultfd_wake(ctx, arg);
1989 break;
1990 case UFFDIO_COPY:
1991 ret = userfaultfd_copy(ctx, arg);
1992 break;
1993 case UFFDIO_ZEROPAGE:
1994 ret = userfaultfd_zeropage(ctx, arg);
1995 break;
1996 case UFFDIO_WRITEPROTECT:
1997 ret = userfaultfd_writeprotect(ctx, arg);
1998 break;
1999 case UFFDIO_CONTINUE:
2000 ret = userfaultfd_continue(ctx, arg);
2001 break;
2002 }
2003 return ret;
2004}
2005
2006#ifdef CONFIG_PROC_FS
2007static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2008{
2009 struct userfaultfd_ctx *ctx = f->private_data;
2010 wait_queue_entry_t *wq;
2011 unsigned long pending = 0, total = 0;
2012
2013 spin_lock_irq(&ctx->fault_pending_wqh.lock);
2014 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2015 pending++;
2016 total++;
2017 }
2018 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2019 total++;
2020 }
2021 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2022
2023 /*
2024 * If more protocols will be added, there will be all shown
2025 * separated by a space. Like this:
2026 * protocols: aa:... bb:...
2027 */
2028 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2029 pending, total, UFFD_API, ctx->features,
2030 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2031}
2032#endif
2033
2034static const struct file_operations userfaultfd_fops = {
2035#ifdef CONFIG_PROC_FS
2036 .show_fdinfo = userfaultfd_show_fdinfo,
2037#endif
2038 .release = userfaultfd_release,
2039 .poll = userfaultfd_poll,
2040 .read = userfaultfd_read,
2041 .unlocked_ioctl = userfaultfd_ioctl,
2042 .compat_ioctl = compat_ptr_ioctl,
2043 .llseek = noop_llseek,
2044};
2045
2046static void init_once_userfaultfd_ctx(void *mem)
2047{
2048 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2049
2050 init_waitqueue_head(&ctx->fault_pending_wqh);
2051 init_waitqueue_head(&ctx->fault_wqh);
2052 init_waitqueue_head(&ctx->event_wqh);
2053 init_waitqueue_head(&ctx->fd_wqh);
2054 seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2055}
2056
2057SYSCALL_DEFINE1(userfaultfd, int, flags)
2058{
2059 struct userfaultfd_ctx *ctx;
2060 int fd;
2061
2062 if (!sysctl_unprivileged_userfaultfd &&
2063 (flags & UFFD_USER_MODE_ONLY) == 0 &&
2064 !capable(CAP_SYS_PTRACE)) {
2065 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2066 "sysctl knob to 1 if kernel faults must be handled "
2067 "without obtaining CAP_SYS_PTRACE capability\n");
2068 return -EPERM;
2069 }
2070
2071 BUG_ON(!current->mm);
2072
2073 /* Check the UFFD_* constants for consistency. */
2074 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2075 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2076 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2077
2078 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2079 return -EINVAL;
2080
2081 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2082 if (!ctx)
2083 return -ENOMEM;
2084
2085 refcount_set(&ctx->refcount, 1);
2086 ctx->flags = flags;
2087 ctx->features = 0;
2088 ctx->state = UFFD_STATE_WAIT_API;
2089 ctx->released = false;
2090 ctx->mmap_changing = false;
2091 ctx->mm = current->mm;
2092 /* prevent the mm struct to be freed */
2093 mmgrab(ctx->mm);
2094
2095 fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2096 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2097 if (fd < 0) {
2098 mmdrop(ctx->mm);
2099 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2100 }
2101 return fd;
2102}
2103
2104static int __init userfaultfd_init(void)
2105{
2106 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2107 sizeof(struct userfaultfd_ctx),
2108 0,
2109 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2110 init_once_userfaultfd_ctx);
2111 return 0;
2112}
2113__initcall(userfaultfd_init);